Abstracts

Rationale: High frequency oscillations (HFOs) have recently received attention as a reliable biomarker of the seizure onset zone. HFOs are often visually identified by expert reviewers, which is a time consuming and subjective process. Automatic detectors show promising results, but they typically require optimization of various parameters and multiple post-processing steps, which vary across datasets and research centers. We have therefore developed an algorithm to identify high frequency ripple events (80-250 Hz) that are statistically different from the background activity. This method has only one parameter related to the sensitivity/specificity of the detector, so it can be applied consistently and broadly.Methods: We first apply a bandpass filter to the intracranial EEG (iEEG), rectify the result, and then find the peak amplitude of each oscillatory cycle. A histogram of these peak values approximately resembles a truncated normal distribution with a long tail due to the presence of transient high amplitude events. An iterative process is then used to estimate the mean and standard deviation of the background probability distribution and calculate an optimum threshold. This threshold is tuned by adjusting the parameter alpha, related to the percentage of allowable false positive events. We define an HFO as six or more consecutive peaks that are statistically significant to ensure that each event contains three or more oscillations that are distinct from the background. We applied the detector to iEEG data from 3 subjects and compared the results to human visual detection. Each dataset contained 8 channels of data for 2 minutes, recorded at 5 kHz. Both depth electrodes and cortical grids were analyzed, and artifacts were excluded before detection.Results: As expected, higher values of alpha resulted in a lower detection threshold and increased sensitivity, while lower values of alpha resulted in higher specificity with lower sensitivity. With alpha = 0.01, the algorithm detected a total of 614 events from the three datasets, and 68% of these events had also been marked by human reviewers. With alpha = 0.005, there were 309 detected events, with 82% marked visually. The alpha parameter is related to the probability that the detected events are significantly different from background activity. Because the algorithm automatically adjusts to the background distribution for each dataset, the results were consistent across datasets and different electrode types.Conclusions: This algorithm provides a measure of confidence in the detection of HFOs by associating oscillatory iEEG events with a probability that they are different from the background activity. By incorporating only one adjustable parameter, related to the percentage of allowable false positive events, it is possible to apply this technique consistently across data from different research centers with different recording equipment. Overall, use of this method may reduce the subjectivity and variability of HFO detection as a tool for the assessment and localization of epileptic activity.